PaperlayerView exact revision on arXiv

Source-linked AI summary

ANTMAN: Adaptive Need Tracking for Multi-Agent Navigation in Large Information Spaces

Jerry Wang, Haibo Jin, Xiaopeng Yuan, Peng Kuang, Haohan Wang

arXiv:2609.33326v1cs.AI

TL;DR

Static partitioning can make multi-agent coordination grow with information-space segmentation rather than query needs. ANTMAN uses a revisable Need Graph to direct workers as unresolved needs evolve. Across evaluated settings it remains effective, while under 16× information-space growth active coordination increases only 1.23×, compared with more than 15× for partition-driven baselines.

  • Problem

    Partition-based multi-agent systems can make coordination grow with the information space even when a query’s unresolved needs remain fixed.

    Source: §1 · p. 1

  • Method

    ANTMAN maintains a revisable Need Graph of unresolved needs, evidence, attempts, and progress to control worker activation, routing, and task-local recovery.

    Source: §1 · p. 2

  • Results

    Across multi-document QA, controlled information-space scaling, and realistic structured navigation, ANTMAN remains effective; under 16× information-space growth, active coordination rises 1.23× versus more than 15× for LongAgent and CoA.

    Source: §1 · p. 2, §4.2.1 · p. 7, §5 · p. 9

  • Takeaways & Limitations

    Need-conditioned coordination transfers across distinct information substrates without substrate-specific redesign and remains effective with substantially smaller worker models.

    Source: §1 · p. 2

Abstract

from arXiv · show

Information-seeking agents increasingly operate over information spaces that are too large to process exhaustively. Yet many multi-agent systems organize computation around static partitions of the available space, causing coordination to grow with how information is segmented rather than with what the query still requires. We introduce ANTMAN, an adaptive coordination framework that treats evolving unresolved information needs as the unit of runtime coordination. ANTMAN maintains a revisable Need Graph that tracks unresolved requirements, accumulated evidence, prior attempts, and search progress, and uses this state to control worker selection, routing, and task-local recovery as new evidence is discovered. By separating the coordination policy from substrate-specific search interfaces, the same need-conditioned mechanism can operate across different information spaces. Experiments across multi-document question answering, controlled long-context scaling, and realistic structured navigation show that ANTMAN remains effective across settings, including when execution is delegated to substantially smaller worker models. Under a 16x increase in searchable context, ANTMAN increases active coordination by only 1.23x, compared with more than 15x for partition-driven baselines, while preserving strong answer quality.

1 INTRODUCTION

Static partition-based systems make coordination grow with information-space segmentation rather than query demand; ANTMAN instead coordinates around evolving unresolved needs in a revisable Need Graph.

1 INTRODUCTION

  • Under a 16× increase in information-space size, active coordination grows 1.23× for ANTMAN versus more than 15× for partition-driven baselines.The passage also reports slower growth in model calls and inference cost.Source: §1 · p. 1, Figure 1 · p. 2
  • Newly encountered evidence can create or refine requirements and make search directions unproductive, so effective coordination needs a revisable account of what remains unresolved.The framework separates its coordination policy from substrate-specific search interfaces.Source: §1 · p. 2
  • ANTMAN maintains a revisable Need Graph that tracks unresolved requirements, evidence, prior attempts, and progress to guide worker activation, routing, and task-local recovery.New evidence can resolve, refine, or introduce needs, redirecting subsequent search.Source: §1 · p. 2
  • ANTMAN outperforms the strongest non-benchmark-specific baselines by 15.3% on RepoProbe, 18.4% on SWE-QA-Pro, and 27.9% on GAIA.With Qwen3-8B for non-orchestrator components, ANTMAN-H retains 89.2%, 93.5%, and 98.3% of full ANTMAN’s performance on those tasks, respectively.Source: §1 · p. 2
  • ANTMAN’s middle-position gap is only +.015, compared with −.181 for direct full-context inference across Early, Middle, and Late evidence placements.This comparison shows reduced sensitivity to evidence position while preserving strong answer quality.Source: §1 · p. 2

2 BACKGROUND AND MOTIVATION

The background reviews adaptive information seeking and multi-agent coordination, distinguishing what information agents access from how they allocate computation.

2 BACKGROUND AND MOTIVATION

  • Prior work spans iterative retrieval, structured navigation, and multi-agent coordination, addressing information access and computation allocation through distinct approaches.ANTMAN is motivated as a link between these directions through unresolved information needs.Source: §2 · p. 3, §2.1 · p. 3, §2.2 · pp. 3–4

2.1 ADAPTIVE INFORMATION SEEKING

  • Adaptive retrieval can update information access as evidence accumulates; S2G-RAG turns insufficient evidence into structured gap queries, while ANTMAN also uses needs to govern worker activation and search effort.ANTMAN extends need-guided information seeking to coordination decisions.Source: §2.1 · p. 3

2.2 MULTI-AGENT NAVIGATION AND COORDINATION

  • Partitioned systems tie coordination footprint to input segmentation, while other systems adapt team composition; ANTMAN uses unresolved needs to connect information access with worker activation, routing, and recovery.Figure 2 presents this workflow through a static substrate map, revisable Need Graph, and need-conditioned coordination.Source: §2.2 · p. 4, Figure 2 · p. 4

3 METHOD

ANTMAN coordinates information seeking around evolving unresolved needs rather than the size of the searchable space. Its Need Graph guides worker activation, routing, and task-local recovery as evidence accumulates.

3.1 PROBLEM FORMULATION

  • ANTMAN makes evolving unresolved information needs the runtime control state, adapting worker activation and search decisions as needs change.The Need Graph tracks unresolved requirements, accumulated evidence, prior attempts, and search progress.Source: §3.1 · p. 4, §3 · p. 4

3.2 STATIC SUBSTRATE MAP

  • The static substrate map partitions the information space into bounded territories, each represented by a scope-and-content WorkerCard.The available WorkerCards define possible workers, not which workers must participate in a query.Source: §3.2 · pp. 4–5

3.3 REVISABLE NEED GRAPH

  • The revisable Need Graph records information needs and dependencies, with each need’s status, accumulated evidence, attempt history, and progress.Worker reports can resolve a need, leave it unresolved, reframe it, or reveal additional requirements and dependencies.Source: §3.3 · p. 5

3.4 NEED-CONDITIONED COORDINATION AND RECOVERY

  • At each step, ANTMAN selects an unresolved need, routes it to a territory-backed worker, and uses the worker’s structured report to update the graph.If attempts stall, task-local recovery can reframe the need, reroute it, or invoke a fallback; workers search only within their assigned scope.Source: §3.4 · p. 5
  • When realized information demand stays bounded, enlarging the number of territories does not force active coordination to grow with the information space.The method bounds coordination steps by the maximum routing attempts per realized need, while active workers also cannot exceed the available territories.Source: §3.4 · p. 5

4 EXPERIMENTS

The experiments assess answer quality in standard multi-document QA, coordination under changing information-space size and demand, transfer to structured navigation, runtime need-state ablations, and evidence-position robustness.

4.1 PRELIMINARY EVALUATION ON STANDARD MULTI-DOCUMENT QA

  • ANTMAN achieves the strongest aggregate performance on standard multi-document QA, while ANTMAN-H remains competitive despite using smaller execution models.ANTMAN performs best on HotpotQA and MuSiQue and remains competitive on 2WikiMultiHopQA; ANTMAN-H matches it on HotpotQA, stays close on 2WikiMultiHopQA, and has a larger gap on MuSiQue.Source: §4.1 · p. 6

4.2.1 CONTROLLED INFORMATION-SPACE SCALING

  • 16× larger searchable context increases active coordination by only 1.23× for ANTMAN, versus over 15× for LongAgent and CoA, while answer quality remains stable.The results also report slower cost growth for ANTMAN as the information space expands.Source: Figure 3 · p. 7, §4.2.1 · p. 7

4.2.2 CONTROLLED INFORMATION-DEMAND SCALING

  • As required evidence increases from 1 to 16 units at fixed 512K space, active coordination rises from 5.0 to 8.0 workers and model calls from 35.5 to 86.2 per query.ANTMAN recovers all required evidence and answers all evaluated queries correctly despite the higher demand.Source: §4.2.2 · p. 7

4.2.3 NAVIGATION IN REALISTIC STRUCTURED INFORMATION SPACES

  • ANTMAN’s need-driven coordination remains effective in repository benchmarks and GAIA-Text-103, with its clearest separation from competing methods on GAIA.ANTMAN-H remains close to full ANTMAN, with the main degradation on harder tasks and most of its GAIA gap arising at Level 3.Source: §4.2.3 · p. 8

4.3 DOES THE EVOLVING NEED STATE MATTER?

  • Full ANTMAN improves answer quality over fixed coordination and graph-free adaptive replanning across all three settings, with recovery and rerouting contributing differently by environment.Recovery is particularly consequential in controlled scaling, while adaptive rerouting matters in repository navigation and has no effect on single-worker GAIA.Source: §4.3 · p. 9

4.4 POSITION ROBUSTNESS UNDER SELECTIVE COORDINATION

  • ANTMAN remains stable across Early, Middle, and Late evidence placements, unlike direct full-context inference, which exhibits strong positional degradation.The results are consistent with selective search narrowing the context considered for each unresolved need.Source: §4.4 · p. 9

5 CONCLUSION

ANTMAN uses a revisable Need Graph to guide runtime coordination as evidence changes. Its execution keeps the territory map fixed while adapting need selection and routing, and bounds active coordination by task demand rather than information-space size.

5 CONCLUSION

  • Across multi-document QA, controlled scaling, and structured navigation, ANTMAN allocates coordination according to evolving needs and remains effective across settings and with smaller worker models.The conclusion also frames this as separating available search capacity from the computation activated for a query.Source: §5 · p. 9

A.1 FULL EXECUTION PROCEDURE

  • The territory map and WorkerCards stay fixed while the Need Graph selects unresolved needs, routes bounded local searches, updates evidence, and triggers recovery when progress stalls.Recovery can reframe a need, reroute it, or use fallback.Source: §A.1 · p. 14
  • Candidate retrieval bounds WorkerCard information considered per routing decision, while the Need Graph governs routing over the query trajectory and termination.The space-decoupling claim concerns active multi-agent coordination, not all retrieval computation.Source: §A.1 · p. 15

A.2 RUNTIME COORDINATION DEMAND

  • Runtime demand tracks unresolved needs as their status changes, includes needs introduced or reframed during execution, and limits each need’s routing attempts under a per-need budget.The realized need set can exceed the initial node set.Source: §A.2 · p. 15

A.3 SPACE-DECOUPLED COORDINATION

ANTMAN bounds active workers by realized task demand and a per-need routing budget rather than directly by the number of available territories. The bound does not imply constant worker use for every query.

A.3 SPACE-DECOUPLED COORDINATION

  • Active workers are bounded by realized task demand and the per-need routing budget, rather than directly by the number of available territories M.Increasing M enlarges the candidate pool but does not itself force additional workers to become active.Source: §A.3 · p. 16
  • Each routing event is associated with a selected need and can activate at most one worker, so the per-need routing limit bounds total worker activations.ANTMAN also has only M territory-backed workers.Source: §A.3 · p. 16
  • When realized information demand is bounded and the per-need routing budget does not depend on M, active coordination need not grow linearly with information-space size.The corollary states this space-decoupling condition explicitly.Source: §A.3 · pp. 16–17
  • The bound is not a constant-worker guarantee for every query: harder tasks may increase workers through more unresolved needs, revisions, or recovery attempts.Such growth is induced by task demand, not simply by adding searchable territories.Source: §A.3 · p. 17

A.4 CONTRAST WITH PARTITION-DRIVEN COORDINATION

Partition-driven coordination activates workers according to information-space partitions, whereas ANTMAN's larger candidate pool remains passive until the runtime Need Graph creates demand. Under comparable information requirements, this avoids the linear coordination growth of a partition-driven strategy.

A.4 CONTRAST WITH PARTITION-DRIVEN COORDINATION

  • A partition-driven rule assigns one worker to each of the M information-space partitions and activates them during initial decomposition.The section contrasts this rule with ANTMAN's demand-driven activation.Source: §A.4 · p. 17
  • ANTMAN's larger candidate worker pool remains passive until the runtime Need Graph creates demand, rather than inheriting partition-driven linear coordination growth.The comparison holds under fixed or comparable information requirements.Source: §A.4 · p. 17
  • The evaluated scaling condition increases searchable-space size while keeping the underlying information requirements comparable.Source: §A.4 · p. 17

B DETAILED CONTROLLED SCALING RESULTS

The detailed controlled-scaling results report endpoint growth and additional per-length and position-sensitivity analyses. Table 7 summarizes how coordination, calls, cost, and answer quality change between 32K and 512K.

B DETAILED CONTROLLED SCALING RESULTS

B.1 ENDPOINT SCALING

  • Table 7 summarizes changes in coordination, model calls, cost, and answer quality from 32K to 512K.Source: §B.1 · p. 17

B.2 COMPLETE PER-LENGTH RESULTS

The complete controlled-scaling results span context lengths from 32K to 512K, with 30 evaluations at each length. The tables describe endpoint growth and the metrics used to assess coordination and answer quality.

B.2 COMPLETE PER-LENGTH RESULTS

  • Table 8 covers a 16× increase in information-space size from 32K to 512K, averaging 30 evaluations per context length.Each context length uses 10 questions at three evidence positions.Source: §B.2 · p. 17
  • Table 7 compares endpoint growth between 32K and 512K and reports the corresponding absolute change in answer quality as ΔF1.Source: Table 7 · p. 18
  • Table 8 reports active coordination, calls, cost, EM, and F1; lower values are better for the first three, while higher values are better for EM and F1.Coordination, calls, and cost are averaged per query.Source: Table 8 · p. 18

C RAW SCORES FOR STRUCTURED NAVIGATION

The structured-navigation appendix reports raw benchmark scores alongside a description of the ablation variants. The variants distinguish persistent Need Graph state from adaptive control mechanisms and are evaluated under matched execution budgets.

C RAW SCORES FOR STRUCTURED NAVIGATION

D.1 VARIANT DEFINITIONS

  • Static ANTMAN freezes the initial Need Graph and routing plan, while the full method retains Need revision, adaptive rerouting, and recovery.Source: §D.1 · p. 19
  • Graph-free Adaptive retains adaptive worker selection and search redirection but has no persistent explicit representation of unresolved needs.It receives the query, WorkerCards, accumulated evidence, and interaction history.Source: §D.1 · p. 19
  • The ablations separately disable Need revision, adaptive rerouting, or recovery; Table 10 summarizes the state and mechanisms available to each variant.Source: §D.1 · p. 19, Table 10 · p. 19
  • All variants are evaluated with matched execution budgets on the 512K controlled condition, SWE-QA-Pro, and a fixed 40-question GAIA subset.Source: §D.1 · p. 19

D.2 COMPLETE RESULTS AND ANALYSIS

Across controlled scaling and repository navigation, the results distinguish explicit need tracking, revision, rerouting, and recovery from simply activating more workers. The runtime behaviors use the Need Graph to guide worker actions and update unresolved needs, while substrate-specific access tools vary across environments.

D.2 COMPLETE RESULTS AND ANALYSIS

  • Explicit Need Graph tracking yields higher controlled-scaling answer quality than Graph-free Adaptive with a comparable number of activated workers.The passage attributes this to persistent unresolved-need tracking directing subsequent search.Source: §D.2 · p. 19, Table 11 · p. 20
  • On SWE-QA-Pro, disabling need revision lowers answer quality despite more worker activations, while disabling adaptive rerouting also reduces answer quality.The results indicate that broader activation alone does not ensure better evidence acquisition and highlight revisiting worker assignments during repository navigation.Source: §D.2 · pp. 19, 21, Table 12 · p. 20
  • Removing recovery lowers controlled-scaling answer quality and, on GAIA, lowers accuracy while increasing model calls and inference cost.In the GAIA setting, recovery is associated with lower realized execution cost as well as higher answer quality.Source: Table 13 · p. 20, §D.2 · p. 21

E PRIMARY RUNTIME BEHAVIORS

  • ANTMAN’s runtime behaviors cover Need Graph construction and revision, need-conditioned worker execution, evidence selection, and need resolution.The behaviors are used during the framework’s runtime coordination loop.Source: E PRIMARY RUNTIME BEHAVIORS · p. 21

E.1 ORCHESTRATOR

  • The orchestrator selects unresolved needs and suitable workers, revises or decomposes needs, respects dependencies, and consolidates proposed needs by adding, linking, merging, subsuming, or discarding them.Consolidation is based on the current graph and proposed need.Source: §E.1 · p. 21

E.2 WORKER EXECUTION

  • For a selected need, a worker uses its assigned territory and substrate-specific tools to plan bounded evidence-seeking actions, then returns evidence, uncertainty, and progress.Source: §E.2 · p. 21

E.3 EVIDENCE AND NEED-STATE UPDATE

  • Evidence selection retains material relevant to resolving the query, and need assessment closes resolved needs while retaining or refining partial and unresolved needs.Partial needs are replaced or refined into more specific unresolved requirements for subsequent coordination.Source: §E.3 · p. 21

F.1 SUBSTRATE-SPECIFIC TOOL INTERFACES

  • ANTMAN keeps its Need Graph and coordination procedure unchanged across substrates while adapting information-access tools, including structural code navigation for repositories and web search for GAIA.Controlled long-context and repository settings share a local access implementation; GAIA uses a separate web-oriented interface.Source: §F.1 · pp. 21–22

F.2 CONTROLLED INFORMATION-DEMAND SCALING DETAILS

In controlled 512K contexts, evidence recovery and answer accuracy remain complete across tested information-demand levels, while worker activity becomes more focused on required evidence as demand rises. RepoProbe-Python diagnostics show that similar interaction counts can accompany substantially different scores.

F.2 CONTROLLED INFORMATION-DEMAND SCALING DETAILS

  • The 30 matched conditions use 10 fixed 512K contexts, three evidence-demand levels, and a shared model, tools, routing, worker organization, and execution budget.Answer quality is exact-match accuracy, and complete evidence requires recovery of all required evidence units from extracted content.Source: §F.2 · p. 22
  • At one required unit, active coordination is 5.0 workers versus 4.3 at four units because low-demand exploration accounts for proportionally more activity.The passage identifies this slight inversion as exploratory overhead rather than failed evidence acquisition.Source: §F.2 · pp. 22–23
  • Complete evidence recovery and answer accuracy remain at 100% across all three demand levels; at one required unit, 4.1 of 5.0 activities are unrelated, while 3.2 of 4.3 contribute at four units.At one required evidence unit, unrelated activity accounts for 82%; at four units, irrelevant activity falls to 1.1.Source: §F.2 · pp. 22–23

F.3 REPOPROBE-PYTHON EXECUTION DIAGNOSTICS

  • 38.21 RepoProbe-Python score for ANTMAN versus 22.87 for S2G-RAG accompanies similar usage: 38.13 versus 37.12 LLM calls and 26.81 versus 24.00 tool calls.All agentic methods have the same maximum of 100 LLM calls per query, and average realized usage is below 40 calls for every method.Source: §F.3 · p. 23, Table 16 · p. 23
  • The sparse-retrieval baseline uses up to three iterative lexical retrieval rounds, with the LLM deciding whether evidence suffices and reformulating queries when needed.Its averages of 2.16 tool calls and 3.19 LLM calls per query reflect early termination on some questions.Source: §F.3 · p. 23

Related papers

Loading 2609.33326v1…